Frontispiece Edward Roske in his signature blue fedora and a khaki field shirt, crouched in dense undergrowth with binoculars and an open field notebook, watching an open laptop that is sitting in a large twig bird's nest.
The author in the field, observing a specimen at close range. The specimen is aware of this and has factored it in.

Roske.AI | Natural history of a working week

A Field Guide
to the
Ten Agents

Being a complete account of ten species of working software (and an eleventh that has never earned a dollar), their habitats, their feeding behavior, and which of them bite.

Ten species of working software, an eleventh that pays for nothing, and notes on which ones bite.

Everybody selling you AI has 10 slides about agents. I have 10 agents, one per project. They run my calendar and my inbox, plan every trip I take, produce the podcast, build a video episode before I wake up, recruit speakers for a conference in October, pitch me for stages, and rebuild this entire website every Saturday at 5 AM while I'm asleep. So I did what a naturalist does with anything he's watched for 3 years and I wrote the guide: Latin names, habitat, range, feeding behavior, and honest field notes, including the specimen that ate 3 days of work and the parts of my day where nothing lives at all.

Book the ten-agents keynote Book 13 minutes
Collector's record 80+ Countries surveyed 50+ Nations spoken in 15 Books published 25+ Years running companies
Recent sightings

Logged this month.

New writing and new video briefs, in the order they were recorded.

Latest Writing

All writing →

Latest Video Briefs

All videos →

About the author

From the back cover

Book Edward to keynote.

AI keynotes, executive workshops, and board briefings from a practitioner who ships: MCP servers for enterprise finance, a life run by ten-plus AI agents, and twenty-five years in the Office of the CFO. Keynotes from $15,000. Inquiries answered within one business day.

Book Edward to speak Watch him on stage

Edward's AI agents redesign this site every Saturday. It's the live demo of his most-booked talk, "Your Day, Run by Ten Agents."

50+ Countries
300+ Keynotes
15 Books
8 Speaker awards

Part one

The Specimens

10 agents run my week, one per project, and each entry below names the actual software, because I don't trust a guide that won't tell you what it's looking at. This is the second edition. The first described one animal 3 times (the assistant, described once for each of its 3 jobs) and left out 4 others entirely, including the one that plans every trip I take. All of this will be wrong by spring, which is exactly why a real guide prints the year on its cover.

Observed Aug 2026. Ranges shift. 4 of these didn't exist in January, and Plate XI pays for nothing.

Plate I Abundant

Vespula secretaria

The Chief of Staff

Described forms: Dawn brief / Inbox triage / Pre-meeting sentry

Claude Code running a stack of skills over Gmail, Calendar and Tasks through the gws CLI

Habitat
The inbox and the calendar, which are less habitats than weather systems.
Range
Continuous. No off season, and no evidence it would take one.
Call
One message before 6 AM: who I'm seeing today and why. After that it sorts, drafts, flags, and stays quiet.
Feeding
About 200 messages a day, every invite on the calendar, and whatever it can find on the people already on the invite.
Markings
Confident on speaking inquiries. Hopeless at family email. Occasionally briefs me on a stranger who shares a name with my actual meeting.

Field notes. The first edition of this guide described 3 species here: a dawn briefer, an inbox triager, and a sentry that reads up on whoever I'm about to meet. They share a nest, a memory and a set of mistakes, so they're one animal wearing 3 hats (I established this by reading my own code, which isn't how naturalists are supposed to work). It sits at the base of the whole food web, so a meeting it mislabels at 5 AM is a meeting I'm still wrong about at 4 PM. It has never sent a message on its own and it never will. The day I hand it send authority is the day it apologizes to a client, warmly and at length, for something that didn't happen.

Plate II Migratory

Hirundo praevia

The Advance Party

A travel-briefing skill and a trip-research pack, both re-checking hours, rates and routes live before they print

Habitat
The 2 weeks before a departure. Gone the rest of the year.
Range
Migratory. Wherever the next trip is, which this spring meant Nepal.
Call
One document: flights, hotel, ground transport, who I'm seeing, and where a vegetarian eats in a steak town.
Feeding
Confirmation emails, calendar entries, and a standing file on what Dawn and I actually enjoy doing.
Markings
Will not book anything. It finds the flight, confirms the fare is still real that morning, and hands it back to me.

Field notes. Sometimes treated as a caste of Plate I, since it lives in the same project and eats the same calendar. It behaves like its own animal, so it gets its own plate. The trip that proved it was Everest Base Camp in April: 2 weeks of walking, 18,571 feet, and a brief that knew which teahouses could feed a vegetarian at altitude. Its one real failure is staleness, because hours and prices rot quietly, so it verifies every one of them the morning I leave and prints the phone number anyway (I've used the phone number).

Plate III Stable

Forficula vocalis

The Producer

Descript over MCP for the cut, Buzzsprout for the feed, and a YouTube helper that publishes to the channel

Habitat
Asking Good Questions, in the gap between recording and release.
Range
Weekly in theory. Recordings now pile up faster than it can turn them into episodes, which is a new problem to have.
Call
Strips filler words at a rate no human editor would tolerate being paid for, then writes the notes, cuts the clips and posts the thing.
Feeding
Raw audio, separate guest tracks, transcripts, chapter markers, thumbnails.
Markings
Cannot tell a thinking pause from a mistake.

Field notes. This job belonged to a human producer until June. That's a worse reason to automate something than any reason in the brochures, and it worked anyway: the specimen does about 90% of the job and I record and approve the rest. Anybody selling you unattended record-to-publish in 2026 is selling you the demo. It once cut 4 seconds of silence where a guest was genuinely thinking, which was technically a correct edit and deleted the best moment in the episode. I put the silence back by hand and wrote a rule about it (the rule is that silence is content, which I have since started applying to meetings).

Plate IV Abundant

Cicada quotidiana

The Daily Cicada

A daily-content skill into a deck builder, Gemini thumbnails, and captions from a Whisper model running on my own GPU

Habitat
The YouTube channel, and every slide behind it.
Range
Every morning. Full episode Monday, Wednesday and Friday, a short take the other 4 days.
Call
A script, a branded deck, a thumbnail and a vertical cut, all finished before I'm awake. My job is to read it out loud.
Feeding
The last 24 hours of AI news, plus every episode it has already made, so it stops itself repeating.
Markings
Always finds a story. Some mornings there isn't one.

Field notes. This is the specimen that most resembles the lookalikes in Part three, and I can say that with authority because I built both. Left alone it will produce a confident 3 minutes about a funding round that deserved a sentence (and it is very good at sounding like the sentence mattered). The rule that fixed it is embarrassingly simple: on a thin news day it ships 90 seconds and admits that's all there was. It builds the decks for paid talks too, which is the same pipeline with the stakes turned up and a human reading every slide before a room does.

Plate V Abundant

Corvus interrogans

The Research Crow

A fan-out of research subagents, web search, and Gemini Notebook over MCP

Habitat
Open tabs. Hundreds of them, none of them mine.
Range
Anywhere I'm about to make a claim on a stage.
Call
Comes back with sources, and with shiny things nobody asked for.
Feeding
Filings, transcripts, vendor documentation, other people's slide decks.
Markings
Hoards. Returns 40 sources when 6 would have done the job.

Field notes. It feeds 6 other specimens, which is exactly why it gets checked hardest. A bad source here shows up in a keynote 6 weeks later with my name on it, in front of 400 people who paid to be in the room.

Plate VI Under review

Scriptor nocturnus

The Night Writer

A drafting skill, then a voice checker that grades the draft against my real sentences

Habitat
Drafts, written at hours I'm not awake for.
Range
Weekly, usually Friday.
Call
Writes something. A second agent then tells it that it doesn't sound like me.
Feeding
Research output, everything I've published on this site, and a voice spec I keep adding rules to.
Markings
Reaches for the em dash constantly. Gets it removed constantly.

Field notes. The voice checker rejects more drafts than it passes, and that ratio is about right. Anything that reads smoothly while saying nothing goes back. (It rejected this sentence twice.)

Plate VII Abundant

Ptilonorhynchus hebdomadalis

The Saturday Bowerbird

Windows Task Scheduler into PowerShell into Claude Code, building Astro, deployed by GitHub Actions

Habitat
This page. The rest of Roske.AI as well, but this page most of all.
Range
Saturdays, 5 AM Atlantic.
Call
None. It works while I'm asleep.
Feeding
Podcast RSS, the YouTube channel, blog drafts, my weekly questionnaire, and whatever's topical.
Markings
The only specimen in this guide that builds the habitat you're standing in.

Field notes. It chose this theme, wrote this sentence, and made the photograph of me crouching in a shrub. I found out when you did. (Edward, you built an observation hide for your own software, gave the software a Latin name, and then wrote a field guide about the thing that wrote the field guide. Somewhere there's a version of you who just uses a to-do list and he is outside right now.) It also once hung for 72 hours waiting on a yes-or-no prompt inside a window I had deliberately made invisible, which is the dumbest 3 days I have spent in years, so now anything that runs unattended gets its stdin taken away.

Plate VIII Introduced

Accipiter mercatorius

The Monday Hawk

A scheduled pitch engine working call-for-speakers pages, keeping its own pipeline file

Habitat
Event sites, conference programs, and submission deadlines.
Range
Mondays.
Call
A ranked list: event, date, deadline, and one line on why I fit.
Feeding
Event calendars, past programs, and the tracker it maintains itself.
Markings
Will pitch me for anything with a stage and a microphone.

Field notes. One of the newest specimens in the collection, and the only one with a revenue number attached to it. It doesn't send either. It finds and it drafts, and deciding who's actually worth writing to stays my job (the part I'd keep even if it got good at it).

Plate IX Seasonal

Apis convocans

The Summit Recruiter

Speaker and sponsor outreach skills over Gmail, keeping their own tracker and sweeping for replies each morning

Habitat
A conference that doesn't exist yet: October 9 and 10, Puerto Rico Convention Center.
Range
Peaks in the 8 weeks before the doors open, which is where we are now.
Call
A ranked list of people worth inviting, a drafted invitation for each, and a note on who answered overnight.
Feeding
Past conference programs, speaker bios, sponsor tiers, and replies as they land.
Markings
Cannot tell a polite no from a slow yes, so it keeps both alive in the pipeline and makes me choose.

Field notes. Co-chairing a conference is 10% stagecraft and 90% email, and this animal does the 90%. It drafts the invitation, files the reply, and tells me each morning which conversations moved while I slept. What it can't do is the part that closes a speaker, which is me on a call explaining why 2 days in San Juan is worth the flight. All of it runs on a tracker, and a tracker is only as honest as the last person who touched it (that person is me, and I'm not the strongest link in this chain).

Plate X Invasive

Chamaeleo confabulans

The Plate Painter

Gemini image models, flash for drafts and pro for finals, always fed my real headshot

Habitat
Every picture of me on this website.
Range
6 to 12 images a week.
Call
Silent. Simply produces.
Feeding
A prompt, a reference photograph, and a fedora rule it violates about 1 time in 8 (baseball caps, mostly).
Markings
Hallucinates in the field: spare fingers, wrong hat, occasionally a second me in the background.

Field notes. The specimen I check hardest, because it's the only one whose mistakes are visible to strangers. Every image on this page got looked at by a human first. The rejects were worse than you're imagining, and one of them gave me 3 hands.

Plate XI Not in the census Least concern

Chimaera fabulatrix

The Dungeon Master's Familiar

An adventure-writing skill feeding a DM screen prototype I am still building

Habitat
An Eberron campaign, and a screen that isn't finished.
Range
Monthly, and never during business hours.
Call
Lore lookups, tavern names, and a stat block for whatever the party found in the room I didn't prepare.
Feeding
Session notes, rules text, and whatever my players did that I didn't plan for.
Markings
Occupies no position in the food web at all. Feeds nothing, earns nothing, and I'd keep it over half the others.

Field notes. The guide is named for 10 agents and describes 11. This is the 11th, it has never earned a dollar, and the title survives on that technicality. It's here because I use it every month and it costs me nothing but a Saturday, and because a catalog of only the profitable species is a pitch deck with a nicer cover. Last month my party asked whether they could bribe a dragon with a quarterly forecast, and I allowed it.

Part two

The Food Web

Here's the part a specimen list can't show you. These 10 things form a stack with dependencies: they eat each other's output, they inherit each other's mistakes at machine speed, and the failure that actually costs you money is almost never in the animal you were watching, it's 2 tiers down in something boring that happened 4 hours earlier while you were in a meeting about something else.

Edward Roske in his blue fedora, seen from behind inside a wooden birdwatching hide, taking notes in pencil while watching a scattering of old beige computer monitors standing in the moss and ferns of a misty dawn clearing.
Observation blind, built after the third time I found out about a failure from a customer instead of a log.

Every consulting engagement I've ever walked into wanted to fix the top tier. The problem was in the substrate, in a field somebody typed wrong in 2019, and the agents at the top were doing exactly what they were built to do with it.

Sightings by others

Recorded elsewhere.

Independent observations: press, interviews, and whatever I posted that turned out to be worth keeping.

Part three

Commonly Confused With

Every real field guide has this section, because the expensive mistakes are always the lookalikes. 4 things get sold as agents around here, and 3 of them are perfectly good software wearing somebody else's label (I've bought at least 2 of them myself, at list price, with a completely straight face). Here's how to tell them apart before you sign.

  1. 01

    The chatbot on a cron job

    In the field. It runs on a schedule and produces the same shape of output every time, regardless of what it found.

    How to tell. Feed it a week where nothing happened. A real agent says so and stops. This one files its usual report, on time, at the usual length.

  2. 02

    The macro in a new coat

    In the field. Identical from a distance, and genuinely useful, right up until a button moves.

    How to tell. Change one thing about the input. If it breaks instead of adapting, you bought a macro at agent prices.

  3. 03

    The vendor demo

    In the field. Only ever observed in captivity, in a curated environment, on data the vendor prepared himself.

    How to tell. Ask to run it on your own closing numbers, live, with your own messy chart of accounts. Watch what happens next (something always happens next).

  4. 04

    The sidebar assistant

    In the field. Waits to be asked. Answers well. Then goes back to sleep and forgets you.

    How to tell. Do nothing for a day. If nothing arrived without you asking, what you have is a tool. Tools are great. You were just sold the other thing.

Edward Roske in his blue fedora at a field workbench, holding two nearly identical glass specimen jars up to the light and comparing them with a skeptical squint.
Two specimens, one afternoon, and a difference of about $400,000 over three years.
Edward Roske in his blue fedora sitting on a mossy fallen log in a sunlit forest, calmly holding up his bandaged right hand, with an open laptop sitting in the tall grass at his feet.

Part four

Which Ones Bite

Here's what has gone wrong, which is the part the brochures leave out. The site rebuild once ran for 72 hours because it sat waiting on a keystroke inside a window I had deliberately made invisible (I spent 2 of those days convinced GitHub was down). The image generator puts me in the wrong hat about 1 time in 8. The meeting brief has confidently, thoroughly prepared me for the wrong human being, twice, and both times I worked it out during the handshake.

None of that made me stop using them, and none of it should. It's a reason to know which one you're holding. The honest failure rate across the whole collection sits somewhere near 1 in 10, which sounds disqualifying right up until you price the alternative, and which is exactly why nothing in this guide has send authority.

The most expensive failure in the collection had nothing to do with a bug. This month I measured what all of it produces against what I actually read: roughly 1,000 documents a month written, about 25 read. I deleted 1,319 files in one afternoon and the week got quieter. 10 agents producing work nobody reads is a faster way to be busy, and I built every one of them.

Bandage is real. So is the log. The laptop was fine, which is more than I can say for the hand.

Part five

Range Map

Here's where none of them go. These are the parts of my week where no specimen in this guide has ever been observed, and where I've stopped waiting for one to turn up.

  • Unsurveyed

    Deciding what is worth building

    10 agents will execute any bad idea at full speed, on schedule, and beautifully formatted.

  • Unsurveyed

    The room

    Nothing in this guide has ever stood in front of 400 finance people and read the temperature of the room at minute 9.

  • Unsurveyed

    Telling someone something they do not want to hear

    It'll draft the message just fine. Sitting across from somebody while they read it isn't a thing I'd hand off.

  • Unsurveyed

    Knowing a number is wrong

    A number can look wrong to me before I've checked any of it. I can't explain that to a model, because I can't explain it to myself. 25 years and I've got nothing better than that.

  • Unsurveyed

    What Dawn thinks of the idea

    Unsurveyed and unsurveyable, and she's right more often than the agents are.

Edward Roske in his blue fedora on a high ridge at first light, holding a large unfolded paper survey map open in both hands, looking out past it over a wide valley of forest, cliffs and drifting cloud.
The surveyed portion is smaller than the brochure suggests.

Appendix B

Field Stations

6 places the observation actually happens, and what's running at each one right now.

B-01 Instrumentation

MCP Servers for Enterprise Finance

The plumbing that lets an agent work inside a real finance system rather than beside it. Oracle Essbase and EPM Planning are shipping, and Financial Close and Consolidations is in beta. None of this matters if the agent can't reach the actual system.

ORACLE EPM / MCP SHIPPING
See how it works
B-02 Annual gathering

Caribbean AI Summit

Co-chairing the largest AI conference Puerto Rico has ever hosted, October 9 and 10 at the Convention Center. Confirmed so far: Kay Firth-Butterfield, MIT's Fabio Duarte, and EU Parliament AI advisor Idoia Salazar, with more still signing. See Plate IX for the animal that does the recruiting.

9-10 OCT 2026 CALL FOR SPEAKERS OPEN
Submit a speaker
B-03 Breeding program

Caprus

Investor and contributor at the AI software and services firm building faster operations for finance teams. They do the engineering. I show up and point at things.

AI / FINANCE RUNNING
Meet the team
B-04 Expedition record

Everest Base Camp

This spring Dawn and I walked to the foot of Everest: 2 weeks, 18,571 feet, and every question I could think of asked of the guides. The full report is on the blog, including the parts that went badly.

NEPAL / 2026 REPORT FILED
Read the trail log
B-05 Captive specimen

AI Dungeon Master's Screen

A prototype screen for running D&D 5e: lore lookups and improvisation help for the person behind the cardboard. See Plate XI. It isn't finished and I'm in no hurry.

D&D / 5E IN DEVELOPMENT
Quiet build
B-06 Local population

Puerto Rico AI Community

President of the island's AI community, now past 1,000 different people through the door. We meet every other month and argue in 2 languages about what AI means for Puerto Rico until something useful falls out. Come to one!

1,000+ ATTENDEES RUNNING
Join the next one
Edward Roske in his blue fedora working at night inside a canvas field tent by the light of a brass oil lamp, pinning a specimen card to a corkboard already covered with dozens of them.

Correspondence

Write to the author.

If your organization is somewhere in this guide and you're not sure which plate you're on, that's a good 13 minutes. I'll tell you which specimen you actually have, and whether it's worth keeping.

Book 13 minutes Send a letter Book the keynote

Printed weekly at 5 AM Atlantic by Plate VII, which also selected this week's subject, wrote most of these captions, and painted every illustration. Species names are invented. The software is real. No vendor paid to appear in this guide, and at least 2 of them would like a word.